def val(net):
# Initialize global and geometric errors ...
num_samples = len(occ_list)
rms = np.zeros(num_samples, np.float32)
log10 = np.zeros(num_samples, np.float32)
abs_rel = np.zeros(num_samples, np.float32)
sq_rel = np.zeros(num_samples, np.float32)
thr1 = np.zeros(num_samples, np.float32)
thr2 = np.zeros(num_samples, np.float32)
thr3 = np.zeros(num_samples, np.float32)
dbe_acc = np.zeros(num_samples, np.float32)
dbe_com = np.zeros(num_samples, np.float32)
dde_0 = np.zeros(num_samples, np.float32)
dde_m = np.zeros(num_samples, np.float32)
dde_p = np.zeros(num_samples, np.float32)
net.eval()
with torch.no_grad():
for i in range(len(occ_list)):
depth_coarse = pred_depths[i].unsqueeze(0).cuda()
occlusion = np.load(os.path.join(opt.occ_dir, occ_list[i]))
# remove predictions with small score
mask = occlusion[:, :, 0] <= opt.th
occlusion[mask, 1:] = 0
occlusion = padding_array(occlusion)
occlusion = occlusion.unsqueeze(0).cuda()
# forward pass
if opt.use_normal:
aux = cv2.imread(os.path.join(opt.data_dir, normal_list[i]), -1) / (2 ** 16 - 1) * 2 - 1
elif opt.use_img:
aux = cv2.imread(os.path.join(opt.data_dir, img_list[i]), -1) / 255
else:
aux = None
if aux is not None:
aux = padding_array(aux).unsqueeze(0).cuda()
if opt.use_log:
depth_refined = depth_coarse * net(depth_coarse.log(), occlusion, aux).exp()
else:
depth_refined = net(depth_coarse, occlusion, aux)
pred = depth_refined.clamp(1e-9)
# get numpy array from torch tensor and crop
gt = gt_depths[i, eigen_crop[0]:eigen_crop[1], eigen_crop[2]:eigen_crop[3]]
edge = gt_boundaries[i, eigen_crop[0]:eigen_crop[1], eigen_crop[2]:eigen_crop[3]]
pred = pred.squeeze().cpu().numpy()[eigen_crop[0]:eigen_crop[1], eigen_crop[2]:eigen_crop[3]]
gt_vec = gt.flatten()
pred_vec = pred.flatten()
abs_rel[i], sq_rel[i], rms[i], log10[i], thr1[i], thr2[i], thr3[i] = compute_global_errors(gt_vec, pred_vec)
dbe_acc[i], dbe_com[i], est_edges = compute_depth_boundary_error(edge, pred)
dde_0[i], dde_m[i], dde_p[i] = compute_directed_depth_error(gt_vec, pred_vec, 3.0)
return abs_rel, sq_rel, rms, log10, thr1, thr2, thr3, dbe_acc, dbe_com, dde_0, dde_m, dde_p
np.ascontiguousarray(depths)).float().unsqueeze(1)
assert len(occ_list) == depths.shape[
0], 'depth map and occlusion map does not match !'
with torch.no_grad():
for i in tqdm(range(len(occ_list)),
desc='refining depth prediction from {}'.format(method)):
depth_coarse = depths[i].unsqueeze(0).cuda()
occlusion = np.load(os.path.join(opt.occ_dir, occ_list[i]))
# remove predictions with small score
mask = occlusion[:, :, 0] <= opt.th
occlusion[mask, 1:] = 0
occlusion = padding_array(occlusion)
occlusion = occlusion.unsqueeze(0).cuda()
# forward pass
if opt.use_normal:
aux = cv2.imread(os.path.join(opt.data_dir, normal_list[i]),
-1) / (2**16 - 1) * 2 - 1
elif opt.use_img:
aux = cv2.imread(os.path.join(opt.data_dir, img_list[i]),
-1) / 255
else:
aux = None
if aux is not None:
aux = padding_array(aux).unsqueeze(0).cuda()
pred = net(depth_coarse, occlusion, aux)